Use stddev from benchmark to do a statistical test

[fonsp]

We currently get some random test failures, because the benchmarks are hardcoded to be "within 1.2 times the time from Distributed". I found that:

• @belapsed is the minimum time, not mean time. That's still an interesting measure (which gets more accurate with more samples) but I believe it has a higher spread? In any case, my intention is to test mean time.
• seconds=1 makes our tests fast but not worth the high random failure rate
• By doing actual stats, we can control the "admissible false positive" rate, which I set to 2.5% right now. It should be pretty low because we have so many CI runners.

Right now the PR changes our goal to be "97% sure that we are not slower".
This should probably be: "97% sure that we are not more than 20% slower". EDIT no its fine

[fonsp]

Hm, in the BenchmarkTools docs there is a section about comparing results, but they just use a fixed tolerance of 1.05 (customizable), instead of taking the stddev into account...

https://juliaci.github.io/BenchmarkTools.jl/stable/manual/#Handling-benchmark-results

They also recommend median instead of mean to be less sensitive to outliers. That makes sense but I forgot how to do statistics with medians instead of means...

[fonsp]

leaving the windows failure for another day :)